AI-moderated HCP interviews built for drug-safety review.

Code decides what to ask and checks every question before it is sent. Rules flag safety mentions in guide answers against EPHMRA criteria and date each from the stored answer, and a person decides every flag. Every finding cites the transcript turn it came from.

Moderation you can audit

A deterministic engine runs the discussion guide. It tracks which must-capture items each answer covered, follows up a vague answer with a question about one specific patient, steers a drift back once, and decides every next move. A language model only phrases the question, inside a reserved live segment of at most 10 questions per interview, and the engine’s own draft is sent whenever the model is slow, refused or breaks a rule.

Recorded interviews: 0 rewordings sent. Safety follow-up questions always use fixed wording, never AI.

Each moderator turn records why it was asked and what the gate decided, and a scorecard grades every interview on process checks (from the recorded moves) and outcome checks (from the respondent’s own words), against design targets that are labelled as targets.

Pharmacovigilance built in

Rules flag adverse events, special reporting situations and product complaints in every answer to a guide question, against the EPHMRA 2025 minimum criteria and each country’s special-situation list. Answers to the safety follow-up and stop questions are not scanned; a person reads them. A flag is a stored record carrying the detector version that raised it; its forwarding deadline, one business day in the interview country’s calendar, counts from the moment the answer was stored, even when the mention is found later.

Recorded example: due Mon 28 Sep, 16:23 ET (synthetic time).

The respondent is never interrupted: follow-up questions for each case come at the end of the interview. There are none after a researcher’s Stop now or End now, or for a mention found after the interview. A person decides every flag (simulated here), dismissing one that meets the criteria takes a second decision, and a reviewer must read every transcript, not only the flags. Each flag exports as an AER-shaped form, and the reconciliation lists every fielded country, zeros included.

When an answer is flagged, the engine does not follow up on it, and the gate blocks any question that names or quotes it, until the guide ends.

Every answer is coded by rules written against a codebook and, where a model run is recorded or live, by a model as a second coder. A validator drops any model quote that is not word for word in the turn it cites, and says so in a count. A theme is counted only where the two coders agree or a person adjudicated the disagreement; the model never writes a theme or a number.

Transcripts, consents, flags and review decisions are hash-chained, so an edit to a stored record shows up when the chain is verified. The sponsor gets a one-page topline and a reconciliation without verification data, screener detail or anything about other companies’ products.

The hash chain detects any edit to a stored record. It does not stop someone with database-owner access from rewriting the whole chain; production would anchor chain heads outside the database (not in v1). This is not a 21 CFR Part 11 audit trail.

Sagan is built for structured and semi-structured interviews where coverage, safety handling and traceability matter more than improvisation. Human moderation is better for:

• Exploratory discovery. An open-ended first look at a new space needs a moderator who can follow a thread the guide never anticipated.
• Sensitive or trauma topics. Bereavement, end-of-life care or anything that needs a person's judgement in the moment.
• Ultra-rare populations. When every respondent is irreplaceable, a human moderator is worth the cost.
• Emotion recognition. Not implemented, and off by design: nothing reads tone, face or voice.
• Anything needing clinical judgement. The moderator never advises, assesses safety or interprets a case.
• Voice. This version is text only: a mis-heard brand or molecule name would break the product criterion silently.